Trained On AI training-data disclosure

Which AI models have a published training‑data disclosure?

9 of 35 tracked models do.

Article 53 of the EU AI Act requires providers of general-purpose AI models to publish a summary of their training content. California's AB 2013 asks a similar question of the same companies.

Article 53 of the EU AI Act and California's AB 2013 both require providers to publish what their models were trained on. What that means.

Two deadlines apply. Models on the EU market from 2 August 2025 had to comply at once; those already on sale have until 2 August 2027. For most models no provider states a placement date at all, so “not published” is not the same as late — see the breakdown below. What Article 53 requires · how this is compiled.

35 models tracked · 9 have published a training-data summary · every claim dated and linked to its source

Models and the Article 53 documents their providers have published
Article 53 disclosures
Model Training-data summary Copyright policy Code of Practice Last checked
Microsoft — Phi-4 Published Not published Signed 2026-08-24
OpenAI — ChatGPT Images 2.0 Published Published — 2026-08-24
OpenAI — GPT-5.2 Published Published — 2026-08-24
OpenAI — GPT-5.4 nano Published Published — 2026-08-24
OpenAI — GPT-5.5 Published Published — 2026-08-24
OpenAI — GPT-5.6 Luna Published Published — 2026-08-24
Swiss AI Initiative — Apertus 70B Published Published — 2026-08-24
xAI — Grok 4.5 Published Published Declined 2026-09-01
xAI — Grok Voice Think Fast 2.0 Published Published — 2026-09-01
AI21 Labs — Jamba Not published Not published — 2026-08-24
Aleph Alpha — Pharia-1 Not published Not published Signed 2026-08-24
Alibaba — Qwen3 Not published Not published — 2026-08-24
Allen Institute — OLMo 2 Not published Not published — 2026-08-24
Amazon — Nova Pro Not published Not published Signed 2026-08-24
Anthropic — Claude Sonnet 4.5 Not published Not published Signed 2026-09-01
Cohere — Command A Not published Not published Signed 2026-09-01
DeepSeek — V3 Not published Not published — 2026-08-24
Google — Gemini 2.5 Flash Image Not published Not published Signed 2026-09-01
Google — Gemma 3 Not published Not published Signed 2026-09-01
IBM — Granite 3 Not published Not published Signed 2026-09-01
Meta — Llama 4 Not published Not published — 2026-09-01
Mistral AI — Mistral Large Not published Not published Signed 2026-08-24
NVIDIA — Nemotron Not published Not published — 2026-08-24
OpenAI — GPT-5 Not published Unverified Signed 2026-08-24
OpenAI — gpt-oss-120b Not published Not published Signed 2026-09-01
OpenAI — Sora 2 Not published Unverified Signed 2026-08-24
Stability AI — Stable Diffusion 3.5 Not published Not published — 2026-09-01
TII — Falcon 3 Not published Not published — 2026-09-01
xAI — Grok 3 Not published Not published Partial 2026-08-24
Black Forest Labs — FLUX 3 Unverified Unverified — 2026-09-01
Bria AI — BRIA 3.2 Unverified Unverified Signed 2026-08-24
Google — Gemini 3 Pro Unverified Unverified — 2026-08-24
Hugging Face — SmolLM3-3B Unverified Not published — 2026-08-24
Meta — Muse Spark Unverified Unverified — 2026-08-24
SpeakLeash — Bielik v3 11B Instruct Unverified Unverified — 2026-08-24

Which models are actually obliged to publish today

Which of the two Article 53 deadlines applies decides whether an empty cell means late or not yet due.

11

Obliged from release

Placed on the market after the cutoff.

6

Have until 2027

On sale before the cutoff, so nothing is owed yet — 335 days until 2 August 2027.

18

Cannot be determined

The provider does not state when the model was placed on the market.

One model is past the cutoff with no summary found

Share of the register established, over time

Reconstructed from the commit history — one point per day the register was re-checked. This tracks our verification, not vendor behaviour: no provider published anything new in this window.

82.9% established
2 Aug from 12% to 82.9% of 35 models 1 Sept
View the underlying numbers
DateTracked PublishedNot published UnverifiedEstablished
2026-08-02 25 1 2 22 12%
2026-08-03 25 0 0 25 0%
2026-08-04 25 1 6 18 28%
2026-08-05 25 3 16 6 76%
2026-08-06 25 3 10 12 52%
2026-08-08 25 3 19 3 88%
2026-08-09 25 3 18 4 84%
2026-08-10 25 2 18 5 80%
2026-08-12 25 2 18 5 80%
2026-08-13 32 7 20 5 84.4%
2026-08-16 35 9 20 6 82.9%
2026-08-17 35 9 20 6 82.9%
2026-08-19 35 9 20 6 82.9%
2026-08-24 35 10 20 5 85.7%
2026-08-31 35 10 20 5 85.7%
2026-09-01 35 9 20 6 82.9%
By provider

One model is not a company

Providers with more than one model on the register. Disclosure is recorded per model because the same company often publishes for one and not another — a provider-level row would be true and false at once.

  • OpenAI 5/8
  • Google 0/3
  • xAI LLC 2/2
  • Meta 0/2

2 of 4 multi-model providers have no confirmed summary on this register

By obligation

Which duties get met

Share of tracked models where the document actually exists. The spread is the finding: the cheap, factual duties are met; the ones that require disclosing training content are not.

  • GPAI provider 100%
  • Model documentation 57%
  • Summary URL 29%
  • Training data summary 26%
  • AI Act statement 26%
  • Follows Commission template 23%
  • Copyright policy 23%
  • California AB 2013 disclosure 3%
  • AB 2013 disclosure URL 3%
  • Systemic risk (stated) 0%